pipeline.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import inspect
from dataclasses import dataclass
from typing import Any, Callable, Dict, List, Optional, Union

import numpy as np
import torch
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection

# Updated to use absolute paths
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.models.unet_motion_model import MotionAdapter
from diffusers.schedulers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
)
from diffusers.utils import (
    USE_PEFT_BACKEND,
    BaseOutput,
    logging,
    scale_lora_layers,
    unscale_lora_layers,
)
from diffusers.utils.torch_utils import is_compiled_module, randn_tensor

# Added imports based on the working paths
from diffusers.models import ControlNetModel
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.utils import deprecate

import torchvision
import PIL
import PIL.Image
import math

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import MotionAdapter, AnimateDiffPipeline, DDIMScheduler
        >>> from diffusers.utils import export_to_gif
        >>> adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2")
        >>> pipe = AnimateDiffPipeline.from_pretrained("frankjoshua/toonyou_beta6", motion_adapter=adapter)
        >>> pipe.scheduler = DDIMScheduler(beta_schedule="linear", steps_offset=1, clip_sample=False)
        >>> output = pipe(prompt="A corgi walking in the park")
        >>> frames = output.frames[0]
        >>> export_to_gif(frames, "animation.gif")
        ```
"""


def tensor2vid(video: torch.Tensor, processor, output_type="np"):
    # Based on:
    # https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/pipelines/multi_modal/text_to_video_synthesis_pipeline.py#L78

    batch_size, channels, num_frames, height, width = video.shape
    outputs = []
    for batch_idx in range(batch_size):
        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
        batch_output = processor.postprocess(batch_vid, output_type)

        outputs.append(batch_output)

    return outputs


@dataclass
class AnimateDiffPipelineOutput(BaseOutput):
    frames: Union[torch.Tensor, np.ndarray]


class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin):
    r"""
    Pipeline for text-to-video generation.
    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
    The pipeline also inherits the following loading methods:
        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
        tokenizer (`CLIPTokenizer`):
            A [`~transformers.CLIPTokenizer`] to tokenize text.
        unet ([`UNet2DConditionModel`]):
            A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
        motion_adapter ([`MotionAdapter`]):
            A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
    """

    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
    _optional_components = ["feature_extractor", "image_encoder"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        motion_adapter: MotionAdapter,
        scheduler: Union[
            DDIMScheduler,
            PNDMScheduler,
            LMSDiscreteScheduler,
            EulerDiscreteScheduler,
            EulerAncestralDiscreteScheduler,
            DPMSolverMultistepScheduler,
        ],
        feature_extractor: CLIPImageProcessor = None,
        image_encoder: CLIPVisionModelWithProjection = None,
    ):
        super().__init__()
        unet = UNetMotionModel.from_unet2d(unet, motion_adapter)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            motion_adapter=motion_adapter,
            scheduler=scheduler,
            feature_extractor=feature_extractor,
            image_encoder=image_encoder,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)

    def load_motion_adapter(self,motion_adapter):
        self.register_modules(motion_adapter=motion_adapter)
    
    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
    def encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.
        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            lora_scale (`float`, *optional*):
                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
        """
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if not USE_PEFT_BACKEND:
                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
            else:
                scale_lora_layers(self.text_encoder, lora_scale)

        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        if prompt_embeds is None:
            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                )
                logger.warning(
                    "The following part of your input was truncated because CLIP can only handle sequences up to"
                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = text_inputs.attention_mask.to(device)
            else:
                attention_mask = None

            if clip_skip is None:
                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
                prompt_embeds = prompt_embeds[0]
            else:
                prompt_embeds = self.text_encoder(
                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
                )
                # Access the `hidden_states` first, that contains a tuple of
                # all the hidden states from the encoder layers. Then index into
                # the tuple to access the hidden states from the desired layer.
                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
                # We also need to apply the final LayerNorm here to not mess with the
                # representations. The `last_hidden_states` that we typically use for
                # obtaining the final prompt representations passes through the LayerNorm
                # layer.
                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)

        if self.text_encoder is not None:
            prompt_embeds_dtype = self.text_encoder.dtype
        elif self.unet is not None:
            prompt_embeds_dtype = self.unet.dtype
        else:
            prompt_embeds_dtype = prompt_embeds.dtype

        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

            max_length = prompt_embeds.shape[1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = uncond_input.attention_mask.to(device)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
            # Retrieve the original scale by scaling back the LoRA layers
            unscale_lora_layers(self.text_encoder, lora_scale)

        return prompt_embeds, negative_prompt_embeds

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
        dtype = next(self.image_encoder.parameters()).dtype

        if not isinstance(image, torch.Tensor):
            image = self.feature_extractor(image, return_tensors="pt").pixel_values

        image = image.to(device=device, dtype=dtype)
        if output_hidden_states:
            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_enc_hidden_states = self.image_encoder(
                torch.zeros_like(image), output_hidden_states=True
            ).hidden_states[-2]
            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
                num_images_per_prompt, dim=0
            )
            return image_enc_hidden_states, uncond_image_enc_hidden_states
        else:
            image_embeds = self.image_encoder(image).image_embeds
            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
            uncond_image_embeds = torch.zeros_like(image_embeds)

            return image_embeds, uncond_image_embeds

    # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
    def decode_latents(self, latents):
        latents = 1 / self.vae.config.scaling_factor * latents

        batch_size, channels, num_frames, height, width = latents.shape
        latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)

        image = self.vae.decode(latents).sample
        video = (
            image[None, :]
            .reshape(
                (
                    batch_size,
                    num_frames,
                    -1,
                )
                + image.shape[2:]
            )
            .permute(0, 2, 1, 3, 4)
        )
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        video = video.float()
        return video

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
        The suffixes after the scaling factors represent the stages where they are being applied.
        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
        Args:
            s1 (`float`):
                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            s2 (`float`):
                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
                mitigate "oversmoothing effect" in the enhanced denoising process.
            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
        """
        if not hasattr(self, "unet"):
            raise ValueError("The pipeline must have `unet` for using FreeU.")
        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
    def disable_freeu(self):
        """Disables the FreeU mechanism if enabled."""
        self.unet.disable_freeu()

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
    def check_inputs(
        self,
        prompt,
        height,
        width,
        callback_steps,
        negative_prompt=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )
        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
    def prepare_latents(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents
    
    def prepare_latents_same_start(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None, context_size=16, overlap=4, strength=0.5):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)
            
        # make every (context_size-overlap) frames have the same noise
        loop_size = context_size - overlap
        loop_count = num_frames // loop_size
        for i in range(loop_count):
            # repeat the first frames noise for i*loop_size frame
            # lerp the first frames noise
            latents[:, :, i*loop_size:(i*loop_size)+overlap, :, :] = torch.lerp(latents[:, :, i*loop_size:(i*loop_size)+overlap, :, :], latents[:, :, 0:overlap, :, :], strength)
            
        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    def prepare_latents_consistent(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None,smooth_weight=0.5,smooth_steps=3):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)

            # blend each frame with the surrounding N frames making sure to wrap around at the end
            for i in range(num_frames):
                blended_latent = torch.zeros_like(latents[:, :, i])
                for s in range(-smooth_steps, smooth_steps + 1):
                    if s == 0:
                        continue
                    frame_index = (i + s) % num_frames
                    weight = (smooth_steps - abs(s)) / smooth_steps
                    blended_latent += latents[:, :, frame_index] * weight
                latents[:, :, i] = blended_latent / (2 * smooth_steps)
                
            latents = torch.lerp(randn_tensor(shape, generator=generator, device=device, dtype=dtype),latents, smooth_weight)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    def prepare_motion_latents(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, 
                        latents=None, x_velocity=0, y_velocity=0, scale_velocity=0):
        shape = (
            batch_size,
            num_channels_latents,
            num_frames,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        
        for frame in range(num_frames):
            x_offset = int(frame * x_velocity)  # Convert to int
            y_offset = int(frame * y_velocity)  # Convert to int
            scale_factor = 1 + frame * scale_velocity

            # Apply offsets
            latents[:, :, frame] = torch.roll(latents[:, :, frame], shifts=(x_offset,), dims=3)  # x direction
            latents[:, :, frame] = torch.roll(latents[:, :, frame], shifts=(y_offset,), dims=2)  # y direction

            # Apply scaling - This is a simple approach and might not be ideal for all applications
            if scale_factor != 1:
                scaled_size = (
                    int(latents.shape[3] * scale_factor),
                    int(latents.shape[4] * scale_factor)
                )
                latents[:, :, frame] = torch.nn.functional.interpolate(
                    latents[:, :, frame].unsqueeze(0), size=scaled_size, mode='bilinear', align_corners=False
                ).squeeze(0)

        return latents

    def generate_correlated_noise(self, latents, init_noise_correlation):
        cloned_latents = latents.clone()
        p = init_noise_correlation
        flattened_latents = torch.flatten(cloned_latents)
        noise = torch.randn_like(flattened_latents)
        correlated_noise = flattened_latents * p + math.sqrt(1 - p**2) * noise

        return correlated_noise.reshape(cloned_latents.shape)

    def generate_correlated_latents(self, latents, init_noise_correlation):
        cloned_latents = latents.clone()
        for i in range(1, cloned_latents.shape[2]):
            p = init_noise_correlation
            flattened_latents = torch.flatten(cloned_latents[:, :, i])
            prev_flattened_latents = torch.flatten(cloned_latents[:, :, i - 1])
            correlated_latents = (prev_flattened_latents * p/math.sqrt((1+p**2))+flattened_latents * math.sqrt(1/(1 + p**2)))
            cloned_latents[:, :, i] = correlated_latents.reshape(cloned_latents[:, :, i].shape)

        return cloned_latents

    def generate_correlated_latents_legacy(self, latents, init_noise_correlation):
        cloned_latents = latents.clone()
        for i in range(1, cloned_latents.shape[2]):
            p = init_noise_correlation
            flattened_latents = torch.flatten(cloned_latents[:, :, i])
            prev_flattened_latents = torch.flatten(cloned_latents[:, :, i - 1])
            correlated_latents = (
                prev_flattened_latents * p
                + 
                flattened_latents * math.sqrt(1 - p**2)
            )
            cloned_latents[:, :, i] = correlated_latents.reshape(
                cloned_latents[:, :, i].shape
            )

        return cloned_latents

    def generate_mixed_noise(self, noise, init_noise_correlation):
        shared_noise = torch.randn_like(noise[0, :, 0])
        for b in range(noise.shape[0]):
            for f in range(noise.shape[2]):
                p = init_noise_correlation
                flattened_latents = torch.flatten(noise[b, :, f])
                shared_latents = torch.flatten(shared_noise)
                correlated_latents = (
                    shared_latents * math.sqrt(p**2/(1+p**2)) + 
                    flattened_latents * math.sqrt(1/(1+p**2))
                )
                noise[b, :, f] = correlated_latents.reshape(noise[b, :, f].shape)

        return noise

    def prepare_correlated_latents(
        self,
        init_image,
        init_image_strength,
        init_noise_correlation,
        batch_size,
        num_channels_latents,
        video_length,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        shape = (
            batch_size,
            num_channels_latents,
            video_length,
            height // self.vae_scale_factor,
            width // self.vae_scale_factor,
        )

        if init_image is not None:
            start_image = ((torchvision.transforms.functional.pil_to_tensor(init_image))/ 255 )[:3, :, :].to("cuda").to(dtype).unsqueeze(0)
            start_image = (
                self.vae.encode(start_image.mul(2).sub(1))
                .latent_dist.sample()
                .view(1, 4, height // 8, width // 8)
                * 0.18215
            )
            init_latents = start_image.unsqueeze(2).repeat(1, 1, video_length, 1, 1)
        else:
            init_latents = None

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )
        if latents is None:
            rand_device = "cpu" if device.type == "mps" else device
            if isinstance(generator, list):
                shape = shape
                # shape = (1,) + shape[1:]
                # ignore init latents for batch model
                latents = [torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)for i in range(batch_size)]
                latents = torch.cat(latents, dim=0).to(device)
            else:
                if init_latents is not None:
                    offset = int(
                        init_image_strength * (len(self.scheduler.timesteps) - 1)
                    )
                    noise = torch.randn_like(init_latents)
                    noise = self.generate_correlated_latents(noise, init_noise_correlation)

                    # Eric - some black magic here
                    # We should be only adding the noise at timestep[offset], but I noticed that
                    # we get more motion and cooler motion if we add the noise at timestep[offset - 1]
                    # or offset - 2. However, this breaks the fewer timesteps there are, so let's interpolate
                    timesteps = self.scheduler.timesteps
                    average_timestep = None
                    if offset == 0:
                        average_timestep = timesteps[0]
                    elif offset == 1:
                        average_timestep = (
                            timesteps[offset - 1] * (1 - init_image_strength)
                            + timesteps[offset] * init_image_strength
                        )
                    else:
                        average_timestep = timesteps[offset - 1]

                    latents = self.scheduler.add_noise(
                        init_latents, noise, average_timestep.long()
                    )

                    latents = self.scheduler.add_noise(
                        latents, torch.randn_like(init_latents), timesteps[-2]
                    )
                else:
                    latents = torch.randn(
                        shape, generator=generator, device=rand_device, dtype=dtype
                    ).to(device)
                    latents = self.generate_correlated_latents(
                        latents, init_noise_correlation
                    )
        else:
            if latents.shape != shape:
                raise ValueError(
                    f"Unexpected latents shape, got {latents.shape}, expected {shape}"
                )
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        if init_latents is None:
            latents = latents * self.scheduler.init_noise_sigma
        # elif self.unet.trained_initial_frames and init_latents is not None:
        #     # we only want to use this as the first frame
        #     init_latents[:, :, 1:] = torch.zeros_like(init_latents[:, :, 1:])
            
        latents = latents.to(device)
        return latents, init_latents

 
    @torch.no_grad()
    # @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        num_frames: Optional[int] = 16,
        context_size=16,
        overlap=2,
        step=1,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_videos_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        output_type: Optional[str] = "pil",
        output_path: Optional[str] = None,
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: Optional[int] = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        clip_skip: Optional[int] = None,
        x_velocity: Optional[float] = 0,
        y_velocity: Optional[float] = 0,
        scale_velocity: Optional[float] = 0,
        init_image: Optional[PipelineImageInput] = None,
        init_image_strength: Optional[float] = 1.0,
        init_noise_correlation: Optional[float] = 0.0,
        latent_mode: Optional[str] = "normal",
        smooth_weight: Optional[float] = 0.5,
        smooth_steps: Optional[int] = 3,
        initial_context_scale: Optional[float] = 1.0,
    ):
        r"""
        The call function to the pipeline for generation.
        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The height in pixels of the generated video.
            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
                The width in pixels of the generated video.
            num_frames (`int`, *optional*, defaults to 16):
                The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
                amounts to 2 seconds of video.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
                expense of slower inference.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                A higher guidance scale value encourages the model to generate images closely linked to the text
                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
                `(batch_size, num_channel, num_frames, height, width)`.
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
                provided, text embeddings are generated from the `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
                `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
                of a plain tuple.
            callback (`Callable`, *optional*):
                A function that calls every `callback_steps` steps during inference. The function is called with the
                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function is called. If not specified, the callback is called at
                every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
        Examples:
        Returns:
            [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is
                returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
        """
        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        num_videos_per_prompt = 1

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
        )

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
            prompt,
            device,
            num_videos_per_prompt,
            do_classifier_free_guidance,
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
            clip_skip=clip_skip,
        )
        # For classifier free guidance, we need to do two forward passes.
        # Here we concatenate the unconditional and text embeddings into a single batch
        # to avoid doing two forward passes
        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

        if ip_adapter_image is not None:
            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
            image_embeds, negative_image_embeds = self.encode_image(
                ip_adapter_image, device, num_videos_per_prompt, output_hidden_state
            )
            if do_classifier_free_guidance:
                image_embeds = torch.cat([negative_image_embeds, image_embeds])

        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        # round num frames to the nearest multiple of context size - overlap
        num_frames = (num_frames // (context_size - overlap)) * (context_size - overlap)
        print(f"Num frames: {num_frames}")
        
        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        if(latent_mode == "normal"):
            latents = self.prepare_latents(
                batch_size * num_videos_per_prompt,
                num_channels_latents,
                num_frames,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
                latents,
            )
        if(latent_mode == "same_start"):
            latents = self.prepare_latents_same_start(
                batch_size * num_videos_per_prompt,
                num_channels_latents,
                num_frames,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
                latents,
                context_size=context_size,
                overlap=overlap,
                strength=init_image_strength,
            )
        elif(latent_mode == "motion"):
            latents = self.prepare_motion_latents(
                batch_size * num_videos_per_prompt,
                num_channels_latents,
                num_frames,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
                latents,
                x_velocity=x_velocity,
                y_velocity=y_velocity,
                scale_velocity=scale_velocity,
            )
        elif(latent_mode == "correlated"):        
            latents, init_latents = self.prepare_correlated_latents(
                init_image,
                init_image_strength,
                init_noise_correlation,
                batch_size,
                num_channels_latents,
                num_frames,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
            )
        elif(latent_mode == "consistent"):
            latents = self.prepare_latents_consistent(
                batch_size * num_videos_per_prompt,
                num_channels_latents,
                num_frames,
                height,
                width,
                prompt_embeds.dtype,
                device,
                generator,
                latents,
                smooth_weight,
                smooth_steps,
            )


        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
        # 7 Add image embeds for IP-Adapter
        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
        
        # divide the initial latents into context groups
        
        def context_scheduler(context_size, overlap, offset, total_frames, total_timesteps):
            # Calculate the number of context groups based on frame count and context size
            number_of_context_groups = (total_frames // (context_size - overlap))
            # Initialize a list to store context indexes for all timesteps
            all_context_indexes = []
            # Iterate over each timestep
            for timestep in range(total_timesteps):
                # Initialize a list to store groups of context indexes for this timestep
                timestep_context_groups = []
                # Iterate over each context group
                for group_index in range(number_of_context_groups):
                    # Initialize a list to store context indexes for this group
                    context_group_indexes = []
                    # Iterate over each index in the context group
                    local_context_size = context_size
                    if timestep <= 1:
                        local_context_size = int(context_size * initial_context_scale)
                    for index in range(local_context_size):
                        # if its the first timestep, spread the indexes out evenly over the full frame range, offset by the group index
                        frame_index = (group_index * (local_context_size - overlap)) + (offset * timestep) + index
                        # If frame index exceeds total frames, wrap around
                        if frame_index >= total_frames:
                            frame_index %= total_frames
                        # Add the frame index to the group's list
                        context_group_indexes.append(frame_index)
                    # Add the group's indexes to the timestep's list
                    timestep_context_groups.append(context_group_indexes)
                # Add the timestep's context groups to the overall list
                all_context_indexes.append(timestep_context_groups)
            return all_context_indexes
                
        context_indexes = context_scheduler(context_size, overlap, step, num_frames, len(timesteps))
        
        # Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        with self.progress_bar(total=len(timesteps)) as progress_bar:
            for i, t in enumerate(timesteps):
                noise_pred_uncond_sum = torch.zeros_like(latents).to(device).to(dtype=torch.float16)
                noise_pred_text_sum = torch.zeros_like(latents).to(device).to(dtype=torch.float16)
                latent_counter = torch.zeros(num_frames).to(device).to(dtype=torch.float16)
        
                # foreach context group seperately denoise the current timestep
                for context_group in range(len(context_indexes[i])):
                    # calculate to current indexes, considering overlap
                    current_context_indexes = context_indexes[i][context_group]

                    # select the relevent context from the latents
                    current_context_latents = latents[:, :, current_context_indexes, :, :]
                    
                    # expand the latents if we are doing classifier free guidance
                    latent_model_input = torch.cat([current_context_latents] * 2) if do_classifier_free_guidance else current_context_latents
                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                    # predict the noise residual
                    noise_pred = self.unet(
                        latent_model_input,
                        t,
                        encoder_hidden_states=prompt_embeds,
                        cross_attention_kwargs=cross_attention_kwargs,
                        added_cond_kwargs=added_cond_kwargs,
                    ).sample

                    # sum the noise predictions for the unconditional and text conditioned noise
                    if do_classifier_free_guidance:
                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)

                        # add the ending frames from noise_pred_uncond to the start of the noise_pred_uncond_sum
                        noise_pred_uncond_sum[:, :,current_context_indexes, :, :] += noise_pred_uncond
                        noise_pred_text_sum[:, :,current_context_indexes, :, :] += noise_pred_text
                        #increase the counter for the ending frames
                        latent_counter[current_context_indexes] += 1

                    # set the step index to the current batch
                    self.scheduler._step_index = i
                
                # perform guidance
                if do_classifier_free_guidance:
                    latent_counter = latent_counter.reshape(1, 1, num_frames, 1, 1)
                    noise_pred_uncond = noise_pred_uncond_sum / latent_counter
                    noise_pred_text = noise_pred_text_sum / latent_counter
                    
                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
                
                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
                
                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()
                    if callback is not None and i % callback_steps == 0:
                        callback(i, t, None)
        
        if output_type == "latent":
            return AnimateDiffPipelineOutput(frames=latents)

        # save frames
        if output_path is not None:
            output_batch_size = 2 # prevents out of memory errors with large videos
            num_digits = output_path.count('#')  # count the number of '#' characters
            frame_format = output_path.replace('#' * num_digits, '{:0' + str(num_digits) + 'd}')
            for batch in range((num_frames + output_batch_size - 1) // output_batch_size):
                start_id = batch * output_batch_size
                end_id = min((batch + 1) * output_batch_size, num_frames)
                video_tensor = self.decode_latents(latents[:, :, start_id:end_id, :, :])
                video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
                for f_id, frame in enumerate(video[0]):
                    frame.save(frame_format.format(start_id + f_id))
            return output_path
            
        # Post-processing
        video_tensor = self.decode_latents(latents)

        if output_type == "pt":
            video = video_tensor
        else:
            video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (video,)

        return AnimateDiffPipelineOutput(frames=video)